Optimization for Determining Production Conditions Using Predictive Models

This study proposes a method for determining optimal production conditions in manufacturing plants. The approach employs a simulator that estimates product quality from production parameters, thereby enabling the search for settings that meet specified quality targets. Reliability is enhanced by utilizing a Structured Neural Network as the surrogate model, which offers interpretable explanations for its predictions. In addition, penalties are applied when production settings deviate from normal operational ranges, guiding the optimization toward conditions that are more realistic for actual operations. Separate simulations carried out with synthetic data and with real production data both demonstrate the method's effectiveness, achieving the desired quality while producing conditions that closely align with real-world operation.